Comparative Study of Software Security Flaw Detection Based on Various Neural Networks

This paper presents a comparative study of various neural networks, including Vanilla LSTM, Stacked LSTM, Bidirectional LSTM, Vanilla GRU, Stacked GRU, Bidirectional GRU, and CNN, to determine the most effective model for software security flaw detection in Python source code. We collected and processed datasets for ten types of vulnerabilities: command injection, cross-origin, directory traversal, open redirect, path disclosure, remote code execution, SQL injection, tampering, XSRF, and XSS. The experiment results clearly indicate that Bidirectional LSTM and GRU models offer the best performance, with high precision and recall across most types of vulnerabilities. Their ability to process sequences in both directions makes them particularly adept at detecting complex and context-dependent security flaws. The results also demonstrate each model's strengths and weaknesses, providing insights into the best practices for automated vulnerability detection in Python codebases.